Skip to content

Quick Start Guide

Fast path from sealed CMP 170HX to an unlocked Linux compute card. Read Prerequisites first so you are not improvising cooling or power mid-flight.

Cooling is step zero

Do not chase unlock scripts on a dry passive heatsink in still air. Confirm airflow or a filled water loop before sustained load. Thermal runaway above ~80 °C is real.

Lab performance headline

Unlocked 64 GiB lab benches lead with Nemotron-3.5-Lightning-30B-A3B W4A16 at 228 tok/s (16k coding), batch 1,306 tok/s aggregate at 32k×16, and 79% decode retained at 128k versus 1k. Earlier Qwen3.8-27B DFlash ~212 tok/s is still in the matrix. Details: Performance Overview.

Setup checklist

Step 1: Identify the SKU

On any Linux host with the card seated (stock driver is fine for ID):

lspci -nn | grep -i nvidia
lspci -s <bus:dev.fn> -vvv | grep -E 'LnkCap|LnkSta|Subsystem'
Device ID Stock label Unlock target
10de:20c2 8 GB 64 GB
10de:2082 10 GB 40 GB

Subsystem checks: 1585 (8 GB) and 1557 (10 GB) per Full Specs.

Difficulty: Easy

Full SKU tables →

Step 2: Host BIOS

Boot the host firmware and set:

  • Above 4G Decoding: Enabled. Needed to map a BAR1 larger than 64 MB. The 64 GB / 40 GB framebuffer size does not depend on it.
  • Secure Boot: Disabled (patched unlock modules are unsigned)
  • Resizable BAR / Re-BAR: Enable if the option is there. Stock BAR1 stays 64 MB until a kernel fixup programs it before PCI enumeration.
  • PCIe slot: Prefer a full x16 mechanical slot even though stock trains at x4

Exact menu names vary by vendor. Capture them on Host BIOS as you learn your board.

Difficulty: Easy–Medium

Step 3: Cooling and power before OS work

  • Seat the 8-pin CPU/EPS-style adapter firmly. This is the stock power feed.
  • Aim a high-static-pressure fan at the passive heatsink, or complete waterblock install first.
  • PSU: plan for 250 W default TDP with headroom to ~300 W software limit, plus host CPU and drives.

Dry-run limit

If you ever power the card without coolant in a waterblock, shut down within about five minutes.

Power → · Air Cooling → · Water Cooling →

Difficulty: Medium (water) / Easy (forced air)

Step 4: Install Linux and nvidia-open 610.43.0x

Community unlock tooling expects:

  • Linux x86-64 (aarch64 appears on some forks; stick to x86-64 unless you know otherwise)
  • nvidia-open 610.43.02 or 610.43.03 already installed (libs + firmware)
  • Kernel headers matching uname -r
  • Network access on first unlock install (sources download)
  • Root privileges

Install your distro, then the matching open driver package set. Freeze or snapshot that driver version before you patch modules.

Difficulty: Medium

Unlock overview →

Step 5: Stock sanity check

Before unlocking, record the stock state:

nvidia-smi
sudo lspci -s <bus:dev.fn> -vvv | grep -E 'LnkSta|LnkCap'

Expect roughly:

  • Memory: 8192 MiB or 10240 MiB
  • Link: Speed 2.5 GT/s, Width x4
  • Idle board power about 30–40 W when the card is cool

If you already see 40–64 GB on a “stock” driver, someone unlocked this host earlier. Document that before you overwrite modules.

Difficulty: Easy

Step 6: Install cmpunlocker

Canonical tool: amoghmunikote/cmpunlocker. Follow that README for the current install.sh flags. Typical profile selection:

sudo ./install.sh --profile=8gb    # 8 GB card → 64 GB geometry
sudo ./install.sh --profile=10gb   # 10 GB card → 40 GB geometry

Then perform a cold reboot (full power off, then boot).

DKMS priority trap (Ubuntu-class hosts)

ServeTheHome forum reports: stock DKMS modules under updates/dkms/ can win over updates/cmpunlocker/. If nvidia-smi still shows 8192 MiB after a “successful” install, check modinfo -n nvidia and add an explicit depmod override pointing at updates/cmpunlocker. Details: Troubleshooting.

Deep mechanism docs: Consensus-Protocol/cmp170hx wiki. Register-level detail stays on that wiki.

Difficulty: Medium–Hard

Step 7: Verify the unlock

nvidia-smi
sudo lspci -s <bus:dev.fn> -vvv | grep LnkSta

Success looks like:

  • Memory: 65536 MiB (8 GB SKU) or 40960 MiB (10 GB SKU)
  • Link speed: 5.0 GT/s when that cmpunlocker build includes Gen2
  • Width: still x4 until you complete the capacitor mod

Optional compute sniff test: a short CUDA or OpenCL FP32 probe should leave the ~0.4 TFLOPS stock FMA class and move toward the ~12–13 TFLOPS unlocked class. Keep the probe short until cooling is proven.

Difficulty: Easy

Step 8: Optional hardware upgrades

Only after unlock verification:

  1. PCIe capacitor mod: 24× 0402 0.22 µF → Gen1 x16 (~4 GB/s). Does not raise Gen by itself.
  2. Water Cooling: Bykski N-TESLA-A100-X-V2 (avoid 80G and non-V2). Pad every empty IC footprint. Air path: Air Cooling.

Difficulty: Hard (SMD) / Hard (loop)

Step 9: Run a real workload

Point an LLM server at the unlocked device. Measured on this card: Nemotron W4A16 228 tok/s (16k coding), 1,306 tok/s aggregate at 16 streams and 32k, 79% of the 1k decode rate at 128k. Tables: Performance.

Quick troubleshooting

Still shows 8 GB / 10 GB after unlock

Check: modinfo -n nvidia path, Secure Boot state, cold vs warm reboot, correct --profile, dmesg for unlock / SEC2 lines.

Check: cmpunlocker build date / branch includes Gen2 (shipped in master since 2026-07-29 per Consensus-Protocol README). Width staying at x4 is expected without capacitor mod.

Thermal throttle or sudden power-off

Check: airflow direction, heatsink seating, water loop fill, nvidia-smi -q -d TEMPERATURE,POWER. Power off if you approach 80 °C without a plan.

Host refuses to POST with card installed

Check: EPS adapter seating, PSU capacity, try another slot, clear CMOS, confirm the card is not shorted after a partial waterblock install.

Troubleshooting →

Next steps

  1. Fill out Prerequisites shopping gaps
  2. Read Specifications once before you publish benches
  3. Follow Unlock Overview for canonical links
  4. Drop numbers into Performance when the lab run is reproducible

References